Protecting Privacy in Federated Time Series Analysis:

A Pragmatic Technology Review for Application Developers*

Daniel Bachlechner

1 a

, Ruben Hetﬂeisch

1 b

, Stephan Krenn

2 c

, Thomas Lorünser

2,3 d

and

Michael Rader

1 e

Fraunhofer Austria Research GmbH, Vienna/Wattens, Austria

AIT Austrian Institute of Technology GmbH, Vienna, Austria

Digital Factory Vorarlberg GmbH, Dornbirn, Austria

ﬂ

Keywords:

Federated Time Series, Privacy-Preserving Technologies, Technology Review.

Abstract:

The federated analysis of sensitive time series has huge potential in various domains, such as healthcare or

manufacturing. Yet, to fully unlock this potential, requirements imposed by various stakeholders must be

fulﬁlled, regarding, e.g., efﬁciency or trust assumptions. While many of these requirements can be addressed

by deploying advanced secure computation paradigms such as fully homomorphic encryption, certain aspects

require an integration with additional privacy-preserving technologies.

In this work, we perform a qualitative requirements elicitation based on selected real-world use cases. We

match the derived requirements categories against the features and guarantees provided by available technolo-

gies. For each technology, we additionally perform a maturity assessment, including the state of standardiza-

tion and availability on the market. Furthermore, we provide a decision tree supporting application developers

in identifying the most promising candidate technologies as a starting point for further investigation. Finally,

existing gaps are identiﬁed, highlighting research potential to advance the ﬁeld.

1 INTRODUCTION

The proliferation of data-driven technologies has led

to an exponential increase in the collection and anal-

ysis of time series data across multiple domains.

Time series data, characterised by its sequential and

temporal nature, provides critical insights that drive

decision-making processes in sectors such as health-

care, manufacturing and smart infrastructure. How-

ever, the sensitivity of this data often includes per-

sonal, conﬁdential or proprietary information, requir-

ing strict privacy measures.

In healthcare, time series data from patient mon-

itoring systems and electronic health records can re-

https://orcid.org/0000-0001-7726-9065

https://orcid.org/0000-0002-9428-3835

https://orcid.org/0000-0003-2835-9093

https://orcid.org/0000-0002-1829-4882

https://orcid.org/0009-0000-1819-5246

∗

Authors are listed in alphabetical order.

https://www.veriﬁedmarketresearch.com/product/

time-series-analysis-software-market/

veal important trends and treatment outcomes, con-

tributing signiﬁcantly to medical research and patient

care. However, patient privacy must be protected to

comply with regulatory standards such as HIPAA and

GDPR, while still allowing for comprehensive data

analysis. This delicate balance between data utiliza-

tion and privacy is a signiﬁcant challenge.

In the manufacturing sector, often referred to as

Industry 4.0, time-series data collected from sensors

and machines is essential for optimising production

processes, predictive maintenance and improving op-

erational efﬁciency. If this data is compromised, it can

lead to signiﬁcant competitive and operational risks.

It is therefore critical to implement privacy technolo-

gies that can protect proprietary information while fa-

cilitating the analysis required for innovation and im-

provement.

Smart buildings, a specialised area of smart in-

frastructure that integrates IoT devices and sensors

to monitor and control various systems such as light-

ing, heating and security, generate large amounts of

time series data. This data can improve energy efﬁ-

ciency, enhance occupant comfort and ensure safety.

198

Bachlechner, D., Hetﬂeisch, R., Krenn, S., Lorünser, T. and Rader, M.

Protecting Privacy in Federated Time Series Analysis: A Pragmatic Technology Review for Application Developers.

DOI: 10.5220/0013356100003950

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 15th International Conference on Cloud Computing and Services Science (CLOSER 2025), pages 198-210

ISBN: 978-989-758-747-4; ISSN: 2184-5042

However, it also contains sensitive information about

the behaviour and preferences of building occupants,

requiring robust data protection measures to prevent

unauthorised access and misuse.

What all these domains have in common is that

valuable insights can be gained by combining data

coming from different – potentially mutually distrust-

ing – entities. Thus, privacy-preserving technologies

(PPTs) providing formal security guarantees are re-

quired to unlock the potential of the data without com-

promising conﬁdentiality. This is also in line with the

paradigm of data sovereignty, as also pushed forward,

e.g., by the European Data Spaces, guaranteeing users

and organizations full control over their own sensitive

data.

Besides social acceptance, the consideration of

PPTs is also mandated by legal frameworks. For

instance, Article 32 GDPR requires to consider the

available state of the art when selecting measures to

ensure a level of security appropriate to the identiﬁed

risks. In this context, as clariﬁed, e.g., by the Eu-

ropean Union Agency for Cybersecurity (ENISA), a

technology is considered as “state of the art” when

it reaches market maturity, which, among others, de-

pends on its availability and level of standardization.

However, the landscape of PPTs is complex and

rapidly evolving, making it challenging for both prac-

titioners and researchers to navigate. Identifying tech-

nologies that have the necessary properties and are

mature enough to be implemented in real-world sce-

narios remains a daunting task. This difﬁculty is com-

pounded by the diversity of PPTs, ranging from se-

cure multi-party computation to federated learning,

each with its own strengths and limitations. As a re-

sult, there is an urgent need for comprehensive evalu-

ations and guidelines to assist in the selection and use

of the most appropriate technologies for speciﬁc time

series analysis applications.

Our Contribution. This research explores the practi-

cal applicability of PPTs in real-world use cases, iden-

tiﬁes research gaps, proposes a high-level research

agenda and provides practical guidance for both prac-

titioners and researchers aiming to implement PPTs

in sensitive time series analysis.

More precisely, we provide developers of col-

laborative time-series analysis applications with a

comprehensive overview of the available computing

paradigms and complementary additional PPTs. For

each technology, we provide an analysis of its proper-

ties with respect to the most relevant requirement cat-

egories derived from real-world use cases and based

on discussions with relevant players in the respective

https://www.enisa.europa.eu/news/enisa-news/

what-is-state-of-the-art-in-it-security

domains, and support the developer by assessing the

maturity level of each candidate.

Furthermore, we provide practical guidance in the

identiﬁcation of suitable candidate technologies for

given scenarios.

Finally, we identify future research potential and

propose a concise research agenda, to further increase

the real-world applicability of available technologies.

Related Work. For an introduction to time-series

analysis in general, we refer to the academic litera-

ture, e.g., (Brockwell and Davis, 2016).

There are systematic overviews of PPTs and guid-

ance for the selection of such technologies both in

general (Bachlechner et al., 2018) and for differ-

ent application domains, e.g., for health data (Jor-

dan et al., 2022) and production (Sielaff et al., 2023).

A comprehensive overview in the context of ofﬁ-

cial statistics has recently been published by the UN

Statistics Division (United Nations, 2023). Support in

the maturity assessment was offered by ENISA.

However, to the best of our knowledge, no com-

prehensive overview supporting developers in assess-

ing the most promising technologies for speciﬁc ap-

plication domains of (federated) time series exists.

Paper Outline. The paper is structured as follows:

The results of a requirements analysis are presented in

Section 2. Advanced secure computation paradigms

as well as complementary PPTs are introduced in Sec-

tion 3, discussing their properties and maturity. Sec-

tion 4 then discusses the pros and cons of all ap-

proaches and provides support for technology selec-

tion. Section 5 concludes the paper with a research

agenda and practical guidelines for practitioner and

researchers.

2 REQUIREMENT CATEGORIES

In the following we introduce a set of exemplary ap-

plication scenarios in the context of federated time

series analysis. Supported by discussions with stake-

holders from the domains, we performed a qualitative

requirements elicitation for each of these uses.

The use cases were chosen from complementary

domains to highlight the diverse needs and expecta-

tions associated with federated time series. Although

numerous speciﬁc use cases exist within each domain,

the selected examples represent a wide spectrum of

requirements, encompassing a broad range of appli-

cations in federated time series analysis.

https://www.enisa.europa.eu/publications/pets

https://www.enisa.europa.eu/publications/

pets-maturity-tool

Protecting Privacy in Federated Time Series Analysis: A Pragmatic Technology Review for Application Developers

199

Based on this, Section 2.2 then presents a cluster-

ing of the found requirements, showing clear similar-

ities and differences in the needs of our use cases.

2.1 Motivating Use Cases

2.1.1 Domain 1: Medical Research

This use case considers the scenario of multiple hos-

pitals collaborating on joint medical studies. This is

particularly relevant in the context of rare diseases,

where a sufﬁciently large number of patients can only

be monitored and analyzed by involving geographi-

cally distributed medical facilities.

A concrete example is related to post-operative

pain management and the required amount of Fen-

tanyl for treatment, as studied by (Chiang et al.,

2016), based on time series containing time stamps

and dosage. Assuming that not all patients are treated

in the same hospital, calculating the signiﬁcance of

different drug dosages can be challenging. Given that

medical data is classiﬁed as special category data ac-

cording to Article 10 of the GDPR, any computation

or pooling of such information is constrained by strin-

gent regulatory and ethical concerns.

Requirements. To facilitate such studies, it would

be advantageous if data controlled by a hospital never

leaves its facility, not even in encrypted form, as this

still constitutes personally identiﬁable information.

Given that different studies typically involve var-

ious types of data and computations, a highly versa-

tile setup is required. However, efﬁciency is only of

secondary importance, and no real-time requirements

exist. A reasonable IT infrastructure in terms of com-

putational and bandwidth capacity is assumed.

Since hospitals are highly regulated, they are un-

likely to maliciously deviate from protocol speciﬁca-

tions. Nonetheless, because sensitive decisions (e.g.,

regarding treatment strategies) may depend on the

outcomes, auditability would still be beneﬁcial to the

greatest extent possible.

2.1.2 Domain 2: Industry 4.0

Among other goals, Industry 4.0 aims at creating

smart factories where machines, operators, and man-

ufacturers are interconnected through the Internet of

Things (IoT). Machines generate vast amounts of data

which they transmit to the Original Equipment Man-

ufacturer (OEM). The OEM uses this data to train so-

phisticated models aimed, e.g., at predictive mainte-

nance and remote diagnostics. This can signiﬁcantly

enhance operational efﬁciency, reduce downtime, and

lead to cost savings, marking a substantial advantage

over traditional production methods.

However, this continuous stream of data also

brings substantial conﬁdentiality concerns, as it may

contain information about production processes or

proprietary technologies. This makes the OEM a cen-

tral trusted party, from which sensitive information

may leak, e.g., to competitors, leading to direct eco-

nomic harm.

Requirements. What is needed, therefore, are tech-

nologies that enable the federated training of ad-

vanced models, without a central entity that learns any

information about the underlying data. Furthermore,

in the optimal case, the different contributors to such

a model do not need to know about each other, i.e., no

direct communication among them is necessary.

Regarding efﬁciency requirements, no real-time

needs exist for the training phase; however, to avoid

damage to the production machines, real-time analy-

sis of the data is required.

The results of this data analysis step may have im-

mediate consequences on business continuity: in case

of false negatives, unnecessary downtime may occur,

increasing costs and reducing availability. Thus, es-

pecially if inference is done remotely, any decision

leading to a machine shutdown needs to be auditable

by third parties to avoid fraud.

Additionally, competitors may have incentives to

introduce malicious data into their data sets (also

known as data poisoning), aiming to increase the

downtime of their competitors. As traditional coun-

termeasures like data validation become harder in dis-

tributed settings without central authority, it would be

desirable to receive cryptographic proofs that all data

included in the training phase was indeed generated

by a genuine sensor of the OEM.

2.1.3 Domain 3: Smart Buildings

Modern buildings are generating huge amounts of

data, which can be used by property management, op-

erators or manufacturers of certain devices (e.g., heat-

ing, cooling), or owners of the building. However,

while data access is clearly managed for certain areas,

it becomes more complicated when residents require

information about the data of other apartments: for

instance, it may be possible to optimize one’s heating

settings depending on the presence of neighbors, e.g.,

during winter holiday seasons.

Another illustrative example considers the follow-

ing platform: residents provide their energy consump-

tion, and receive back feedback comparing their use

(e.g., in form of a quantile) to that of their peers in

the same building with comparable external condi-

tions (e.g., weather, building insulation). On the one

hand such gamiﬁcation could increase residents’ will-

ingness to reduce their energy consumption; on the

CLOSER 2025 - 15th International Conference on Cloud Computing and Services Science

200

other hand this information could also be used to de-

tect outliers, which might indicate needs for repair.

Requirements. Residents may be reluctant to dis-

close detailed time series to a central authority, and

especially to their peers, but they should rather re-

main in full control over their data. On the other hand,

however, usability is of utmost importance in such

a setting, such that residents should not be required,

e.g., to host compute nodes for federated computa-

tions themselves. Thus, the computation should be

offered as-a-service while guaranteeing privacy.

Assuming that feedback is provided, e.g., on a

weekly or monthly basis, no real-time requirements

exist. However, given the computation is to be carried

out repeatedly, pre-computations may be acceptable

to increase efﬁciency. Also, given the gamiﬁcation

approach, no immediate auditability or public veriﬁa-

bility requirements exist.

To minimize costs, no powerful IT infrastructure

should be required to carry out the computations.

2.2 Requirement Overview

Based on the scenarios described above, we can now

derive a set of generic requirement categories that

describe the unique needs of each application case.

Following our goal of providing guidelines also to

non-experts in the ﬁelds in the speciﬁcation of their

applications and subsequently identifying the poten-

tial cryptographic approaches, these categories are de-

scribed on a qualitative level to ease an initial assess-

ment.

Trust Model. This category describes the acceptable

assumptions that may be made in an instantia-

tion of the use case, thereby also describing data

sovereignty needs. For instance, can data be

processed (even in encrypted form) by a cloud

provider, or must the computation happen in-

house? Is there only a need to protect against acci-

dental data leaks, or do active attacks by malicious

parties need to be taken into consideration?

Scalability. This category subsumes in particular the

frequency of computations to be performed, the

data sizes to be processed in each computation, or

the number of users providing data.

Efﬁciency. This category includes actual efﬁciency

requirements such as bandwidth and computa-

tional limitations, but also maximum execution

times for each computation including real-time re-

quirements.

Flexibility. This category focuses on the computa-

tions to be performed, i.e., whether a single, static

computation is computed periodically on different

input data, or whether highly dynamic computa-

tions are to be supported. It also includes the func-

tion to be evaluated (e.g., Boolean or arithmetic

circuit, multiplicative depth of circuits, etc.).

Data Integrity. This category is concerned with the

quality of the processed data. In particular, it

speciﬁes whether input data needs to be checked

for integrity and authenticity, which depends on

the decisions to be taken based on the results, and

on whether data providers are mutually trusting

each other regarding the quality of the their data.

Additionally, this category covers completeness,

ensuring that no data is maliciously ignored dur-

ing computation.

Veriﬁability. Closely related to the previous cate-

gory, this covers all aspects related to the quality

of the computation result. That is, is it necessary

that the receiver of a computation result obtains

formal proofs that the computation was done cor-

rectly, or is the data processor trustworthy in the

sense that is follows the protocol speciﬁcation?

Also, speciﬁc use cases may require public ver-

iﬁability (or auditability), where also a third party

is able to check the validity of the result.

Ease of Integration. This category covers all as-

pects related to available hardware and software

systems, acceptable costs, or the availability of

skilled personnel for design and maintenance of

a system.

Regulatory Constraints. Finally, this class captures

all known legal requirements, e.g., related to data

protection or critical infrastructures that need to

be fulﬁlled by any deployed system.

A qualitative mapping of these requirements to

our use cases can also be found in Table 1.

3 TECHNOLOGIES

In Section 3.1 we will now brieﬂy introduce the dif-

ferent approaches found in the literature for federated

computing in general and for time series in particu-

lar. In Section 3.2 we discuss additional PPTs that

in combination with these paradigms can be used to

overcome existing limitations.

3.1 Secure Computing Paradigms

For each of the following technologies, we provide a

short description of their properties, as well as indica-

tions regarding its maturity and practical applicability.

Protecting Privacy in Federated Time Series Analysis: A Pragmatic Technology Review for Application Developers

201

Table 1: Comparison of challenges for our selected use cases.

UC1 UC2 UC3

Medical Research Industry 4.0 Smart Buildings

Trust model Data should not leave own facility Central data processor for protected data acceptable

Scalability Speciﬁc computations per study Medium to high frequency of same computation

small data sizes large data sizes medium data size

Efﬁciency No speciﬁc requirements Real-time needs for analysis No speciﬁc requirements

Flexibility Computations might not be

known when generating data

Computations clear upon data creation

Data integrity Only authentic data should be included in computations; source pri-

vacy desirable

No speciﬁc needs

Veriﬁability Veriﬁability of computation results desirable No auditing needs

Ease of integra-

tion

No inherent hardware limitations, skilled personnel may be available Strong limitations regarding de-

ployment, integration, costs

Regulatory con-

traints

Strict medium to strict (e.g. for critical

infrastructure)

low to medium

3.1.1 Secure Multi-Party Computation

Secure multi-party computation (MPC) was ﬁrst in-

troduced by (Yao, 1982), subsequently generalized by

(Goldreich et al., 1987) and further researched in a

large body of work, cf., e.g., (Evans et al., 2018). It al-

lows parties to collaboratively perform any joint com-

putation on their respective private inputs, without re-

vealing any other information to the other parties than

what is revealed by the predeﬁned partial outputs to

the individual parties.

Properties. The privacy and integrity guarantees of

MPC are based on a non-collusion assumption. That

is, these properties are maintained as long as no more

than a predeﬁned fraction of nodes collude, while no

guarantees can be given if too many nodes collude

maliciously. In particular, distinctions are made de-

pending on whether an honest majority is assumed or

not. Furthermore, different security notions exist, de-

pending on whether malicious nodes still follow the

protocol speciﬁcation (“passive security”) or whether

they may arbitrarily deviate (“active security”). As a

rule of thumb, the efﬁciency of MPC protocols de-

creases as the attacker model becomes stronger, lead-

ing to a delicate balance between achieving the de-

sired levels of security and efﬁciency.

However, despite its versatility, MPC also comes

with several limitations. Speciﬁcally, while any com-

putation on sensitive inputs can be performed using

MPC, the structure of the computation to be per-

formed strongly inﬂuences the efﬁciency of the re-

sulting protocol. For instance, for protocols based on

Shamir’s secret sharing scheme (Shamir, 1979), ad-

ditions are basically for free, while multiplications

require complex, interactive operations among the

nodes, potentially causing signiﬁcant delays depend-

ing on network latency and bandwidth. Therefore, cir-

cuits with low multiplicative depth are preferable in

this approach. Furthermore, the bandwidth require-

ments often heavily depend on the size of the secret

inputs held by the different parties, making applica-

tions with small(er) input sizes more suitable for MPC

than those with larger input sizes.

Finally, depending on the concrete setup, MPC

can be a useful technical control to (partially) over-

come the challenges imposed by data protection reg-

ulations (Helminger and Rechberger, 2022).

Maturity Level. MPC is a highly versatile tool that

has been used to demonstrate a variety of practical

application domains, ranging from secure auctions

(Bogetoft et al., 2009) over industry 4.0 (Lorünser

et al., 2022a) to ﬁnance (da Gama et al., 2022).

Usable frameworks supporting the development

of MPC protocols exist, e.g., MP-SPDZ (Keller,

2020). Furthermore, several standards developing or-

ganizations (SDOs) have already published standards

on MPC, including ISO/IEC (within its ISO/IEC 4922

series) and IEEE 2842-2021.

Guidelines and open challenges for the integration

of MPCs into data infrastructures such as Data Spaces

have recently been investigated by (Siska et al., 2024).

3.1.2 Fully Homomorphic Encryption

While in traditional encryption schemes basically the

only thing that can be done with a ciphertext is to

decrypt it, fully homomorphic encryption (FHE) al-

lows one to perform arbitrary computations on ci-

phertexts in the encrypted domain. Thus, FHE en-

sures that the operations on the ciphertext correspond

CLOSER 2025 - 15th International Conference on Cloud Computing and Services Science

202

to well-deﬁned operations on the underlying plain-

text data. This enables data owners, e.g., to encrypt

their data before sending them to a powerful cloud

service. The cloud provider may then perform com-

putations (e.g., data analysis on sensitive data) on the

ciphertexts without ever having access to the plaintext

data. Eventually, the cloud provider will return the en-

crypted result to the holder of the secret key who only

needs to decrypt the ﬁnal result.

While ﬁrst envisioned much earlier, the concept

was ﬁrst instantiated only in 2009 by (Gentry, 2009).

It has been an active area of research ever since, re-

sulting in signiﬁcant efﬁciency improvements.

Properties. Evaluating functions using FHE is sev-

eral orders of magnitude slower compared to a plain-

text evaluation (Sidorov et al., 2022). In particular,

after a certain number of operations (e.g., multipli-

cations), highly expensive so-called “bootstrapping”

steps need to be performed to guarantee valid de-

cryptions. This is partially overcome by somewhat

homomorphic encryption schemes (Brakerski, 2012),

which support only a small number (e.g., one) of

multiplications but arbitrarily many additions. They

thereby avoid the need for bootstrapping steps while

limiting the expressiveness of the schemes. Thus,

such schemes can signiﬁcantly improve efﬁciency in

speciﬁc applications.

If used to analyze data coming from different data

sources, special attention needs to be paid to the man-

agement of the master secret key msk, as this could be

used not only to decrypt the computation results but

also the encrypted inputs. Thus, in scenarios where

the intended receiver of the computation result (own-

ing msk) must not get access to the individual en-

crypted inputs, it needs to be ensured that the com-

puting server, e.g., does not leak encrypted input data

to the receiver. This can be achieved on an organiza-

tion level, and by enforcing strict access policies. Al-

ternatively, multikey-FHE schemes, ﬁrst introduced

by (López-Alt et al., 2012), are capable of comput-

ing on ciphertexts encrypted under multiple unrelated

keys can be deployed, yet causing higher computa-

tional costs and additional complexity as all individ-

ual secret keys are required for decryption.

Finally, veriﬁability of the performed computa-

tion can be achieved by deploying speciﬁc schemes,

e.g., (Viand et al., 2023).

Maturity Level. FHE is a very ﬂexible technology

that can be used to perform arbitrary computations

on sensitive data, which has been proved in a variety

of scenarios including federated learning (Zhao et al.,

2024) or healthcare (Munjal and Bhatia, 2022).

Over the last years, FHE has made signiﬁcant ad-

vancements regarding practicability, and a range of

open-source frameworks are available to researchers

and developers

5, 6, 7

. Finally, fully homomorphic en-

cryption is subject to ongoing standardization efforts,

e.g., by ISO/IEC (within the ISO/IEC 28033 series),

ITU-T

, and community standardization efforts

. Fi-

nally, in order to overcome efﬁciency limitations, sub-

stantial efforts are currently taken to provide hardware

acceleration for FHE in the near future.

1011

3.1.3 Functional Encryption

Similar to FHE, functional encryption (FE) (Boneh

et al., 2011) offers a more versatile approach than all-

or-nothing access to encrypted data, by allowing the

holder of the master secret key msk to generate partial

decryption keys sk

. Now, upon receiving a ciphertext

that encrypts a message m, the holder of such a par-

tial decryption key cannot recover the message itself,

but only f (m), where f (·) is a function deﬁned by

the holder of the master secret key. If the key is con-

trolled, e.g., by the data subject, ﬁne-grained access

to computations on sensitive data can be provided.

Several extensions to increase functionality or

lower trust assumptions are available in the literature.

Properties. While requiring proper protection of the

master secret key msk to ensure conﬁdentiality of

all data, functional encryption provides an interesting

trust model, as the decryption – and thus the compu-

tation – is directly carried out by the receiving party,

reducing the need for veriﬁability.

On the downside, FE is only meaningful in scenar-

ios where the function to be computed is known in ad-

vance, as the schemes need to support the right class

of functions. In particular, most efﬁcient FE schemes

are focusing on the computation of inner products (in-

cluding, e.g., weighted sums of data items).

Maturity Level. Functional encryption has been

proven useful in different contexts such as, e.g., ma-

chine learning (Panzade et al., 2024). Libraries for

FE are available in different languages through, e.g.,

the FENTEC project

. To the best of our knowledge,

there are no ongoing efforts for standardizing func-

tional encryption.

https://www.zama.ai/

https://github.com/homenc/HElib

https://www.openfhe.org/

https://www.itu.int/ITU-T/workprog/wp_item.aspx?

isn=17999

https://homomorphicencryption.org/

https://www.darpa.mil//news-events/2021-03-08

https://dualitytech.com/partners/intel/

https://github.com/fentec-project

Protecting Privacy in Federated Time Series Analysis: A Pragmatic Technology Review for Application Developers

203

3.1.4 Trusted Execution Environments

Trusted Execution Environments (TEEs) are secure

areas within a processor designed to protect sensi-

tive code and data from unauthorized access or mod-

iﬁcation by the operating system or other applica-

tions (Sabt et al., 2015). TEEs function by providing

isolated execution spaces that operate alongside the

main operating system, ensuring an additional layer

of security for sensitive processes. In practice, TEEs

work by offering a restricted processing environment

where only authorized code can execute, and memory

is encrypted to prevent external inspection or tamper-

ing. This is critical for handling operations such as

cryptographic key management, biometric data pro-

cessing, and secure ﬁnancial transactions (Lind et al.,

2016). When an application needs to perform a secure

operation, it makes a call to the TEE, which processes

the request in isolation, safeguarding the conﬁden-

tiality and integrity of the data being handled. This

makes TEEs vital for enforcing privacy and security

in a broad range of digital applications.

Properties. This is also essential for compliance with

stringent data protection regulations in industries like

ﬁnance and healthcare. Additionally, TEEs help in

mitigating a range of security threats, including mal-

ware and software vulnerabilities. Their ability to

handle secure transactions and authenticate user inter-

actions in a protected manner makes them invaluable

for mobile banking, DRM (Digital Rights Manage-

ment), and personal identiﬁcation applications. Also

the integrity and authenticity of code execution can be

reliably veriﬁed (Chen and Zhang, 2023). Overall,

TEEs strengthen the trustworthiness of devices and

systems, increasing user conﬁdence in digital plat-

forms. TEEs face several challenges that can impact

their broader application and effectiveness. One of

the primary hurdles is hardware dependency, as TEEs

require speciﬁc processor features to function, which

can limit their use to certain devices and platforms,

thus restricting scalability (Geppert et al., 2022). Inte-

grating TEEs into existing systems also poses signiﬁ-

cant challenges due to the need for specialized knowl-

edge to handle secure communication and operation

management between the TEE and the rest of the sys-

tem. This process is further complicated, as any code

loaded into secure enclaves needs to be signed using

tools provided by the TEE manufacturer to ensure se-

curity, making quick adaptations of the code compli-

cated. Moreover, the complexity of developing and

maintaining secure code that runs within TEEs in-

creases the risk of vulnerabilities if not properly man-

aged. Another main point to consider are strong trust

assumptions that need to be put into the trustworthi-

ness of the hardware manufacturer.

Additionally, cross-platform compatibility issues

can arise, complicating the development of applica-

tions that can operate across different types of devices

and operating systems with TEEs.

Finally, the evolving landscape of cybersecurity

threats continuously tests the resilience of TEEs, ne-

cessitating ongoing updates and patches to main-

tain a robust defense. While this is true for any

security-related implementation – e.g., considering

side-channel attacks –, this issue requires further at-

tention due to the required trust in hardware compo-

nents: in case that an attack exploits hardware vulner-

abilities, patches may be more complex and cost in-

tensive than in the case of pure software applications.

Maturity Level. TEEs are provided by several com-

panies, often integrated into their hardware or soft-

ware products, including, e.g., Intel’s Software Guard

Extensions (SGX), AMD Secure Encrypted Virtual-

ization (SEV), ARM TrustZone, or IBM Secure Ser-

vice Container, just to name a few.

The standardization of TEEs has been primarily

driven by the GlobalPlatform organization

3.1.5 Federated Learning

Federated Learning (FL) is a collaborative machine

learning approach, ﬁrst introduced by (McMahan

et al., 2017). It enables model training without cen-

tralizing data, thus preserving the privacy and secu-

rity (Kone

cný et al., 2016) of training data. There-

fore, it allows organizations to jointly develop mod-

els without sharing sensitive data, fostering collabo-

ration and building more generalizable models (Kiss,

2022). Instead of pooling data centrally, each partici-

pant locally trains a model using their private data and

then sends only the model updates (e.g., gradients) to

a central server. This server aggregates the updates

to improve a global model, which is then sent back

to the participants for further training. This cycle re-

peats, enhancing the global model with every step.

To avoid, e.g., model inversion or model inference

attacks, typically a trusted server is assumed for ag-

gregation, but also secure aggregation protocols exist

(Bonawitz et al., 2017).

Properties. Federated learning offers signiﬁcant ad-

vantages, chieﬂy in privacy preservation, as it allows

data to remain on local devices, reducing the risk of

data breaches (Li et al., 2022). Furthermore, the data

security of FL is usually supplemented with differ-

ential privacy to further protect potentially sensitive

training data (Wei et al., 2020). This decentralized ap-

proach also enhances scalability by distributing com-

https://globalplatform.org/

CLOSER 2025 - 15th International Conference on Cloud Computing and Services Science

204

putation across numerous devices, thus avoiding cen-

tralized processing bottlenecks. Additionally, it sup-

ports data diversity, reﬂecting real-world variations,

and improves model robustness. By leveraging mul-

tiple data sources without direct access, it complies

with data protection regulations, making it suitable

for industries like healthcare and ﬁnance where data

sensitivity is paramount. Overall, federated learning

enables more secure and scalable machine learning

deployments (Zhang et al., 2021).

Despite its advantages, FL also faces several lim-

itations. In the case of non-iid data distributions

(data points are not independently and identically dis-

tributed, but may, e.g., be correlated), the ’client-drift’

problem may cause clients to converge to local op-

tima, leading to slower overall convergence and poor

model performance (Moshawrab et al., 2023). An-

other limitation lies in the difﬁculty of determining

hyperparameter values, hindering the global model’s

performance in each update; a proposed solution in-

volves dynamically adjusting hyperparameters using

a log function to address this issue.

Integration of FL into complex systems comes

with a variety of challenges. Varying computational

capabilities of participant devices, the complexity of

establishing reliable communication channels, or the

need to adapt existing systems to handle decentral-

ized data processing need to be addressed. At the

same time, the consistent and secure participation of

joining and exiting nodes in the network must be ad-

dressed (Moshawrab et al., 2023).

Maturity Level. Federated learning is a rapidly ad-

vancing technique, which has found practical appli-

cations in diverse domains such as healthcare (Muazu

et al., 2024) or ﬁnance (Shi et al., 2023).

Usable frameworks and platforms, such as Tensor-

Flow Federated

and PySyft

(Ziller et al., 2021),

support the development and deployment of feder-

ated learning protocols. Finally, several SDOs are ac-

tively working on federated learning standards, such

as IEEE P3652.1.

3.2 Extensions

In the following, we give a concise overview of pos-

sible extensions that may be integrated to achieve all

the requirements identiﬁed in Table 1.

3.2.1 Advanced Signature Schemes

As described in Section 2, data integrity and authen-

ticity are crucial, especially for sensitive decisions. In

https://www.tensorﬂow.org/federated

https://github.com/OpenMined/PySyft

the medical domain, when data is patient-generated,

it is important to ensure only authentic data from gen-

uine devices is used, reducing the risk of manipula-

tion. For example, a diabetes study should only rely

on data from veriﬁed glucometers.

However, signing data directly poses privacy

risks, as it could link measurements to individuals.

This can be mitigated with group signatures (Chaum

and van Heyst, 1991), where a group manager (e.g.,

a device manufacturer) issues unique signing keys to

users. A user signs data (e.g., encrypted blood sugar

readings), and the veriﬁer (e.g., a hospital) checks au-

thenticity without identifying the speciﬁc user. If nec-

essary, a third party (e.g., an ethics board) can reveal

identities in cases of abuse.

For more ad-hoc scenarios, ring signatures (Rivest

et al., 2001) provide similar privacy protections and

are widely used in cryptocurrencies.

Maturity Level. All the aforementioned technologies

are highly mature and have been included into rele-

vant standards, or are currently subject to standard-

ization, e.g., ISO/IEC 20008. Practical applications

of group signatures include, e.g., direct anonymous

attestation (DAA) implemented in Trusted Platform

Modules (TPMs) as speciﬁed by the Trusted Comput-

ing Group.

3.2.2 Zero-Knowledge Proofs

Most computing paradigms in Section 3.1 lack

built-in veriﬁability. This can be achieved using

zero-knowledge proofs (ZKPs) (Goldwasser et al.,

1985), including compact variants like so-called zk-

SNARKs (Groth, 2016). These protocols allow

a prover to demonstrate knowledge of information

without revealing more than the claim itself.

In MPC, compute nodes could generate a zero-

knowledge proof verifying that input data was (i)

signed by providers, (ii) used in the computation, and

(iii) the output is correct. This has been applied in

smart manufacturing (Lorünser et al., 2022a) and air

trafﬁc management (Lorünser et al., 2022b).

A key application of ZKPs is authentic neural net-

work inference. For instance, a service provider may

want to protect a proprietary and certiﬁed model while

proving to customers it was used for inference. ZKPs

enable this without exposing the model. A step to-

ward practical zero-knowledge inference is zkCNN

(Liu et al., 2021).

Maturity Level. ZKPs have reached a relatively ma-

ture stage, with several practical implementations and

real-world applications, including, e.g., private trans-

actions in blockchains

. Companies like StarkWare,

https://z.cash/

Protecting Privacy in Federated Time Series Analysis: A Pragmatic Technology Review for Application Developers

205

zkSync, and Aztec are leading in practical implemen-

tations of ZKPs. Standardization of zero-knowledge

proofs is advancing, driven, e.g., by community stan-

dardization efforts such as ZKProof.

3.2.3 Advanced Encryption Mechanisms

Consider a system where computations are performed

periodically, and users subscribe to the results. This

creates asynchronous scenarios where eligible users

may not be online at computation time, requiring dy-

namic access control. Traditional methods rely on

a central policy-enforcement point, storing results in

plaintext, which raises data protection concerns. En-

crypting results separately for each user using public-

key encryption is also impractical for large user bases.

A more scalable approach is attribute-based en-

cryption (ABE) (Sahai and Waters, 2005), where

users receive secret keys tied to attributes (e.g., sub-

scription validity). Instead of directly outputting re-

sults, compute nodes encrypt under a central key with

an access policy (e.g., creation date). While cipher-

texts are publicly available, only users with matching

keys can decrypt them.

If recipients are known but ofﬂine, or if the data

receiver in an FHE scenario differs from the holder

of msk, proxy re-encryption (Blaze et al., 1998) can

convert ciphertexts for the intended recipient without

exposing plaintexts.

These methods enforce access control crypto-

graphically, ensuring plaintext results remain inacces-

sible to unauthorized users.

Maturity Level. The maturity of ABE is increasing,

with growing interest and several practical implemen-

tations, but not yet widespread adoption. Yet, several

companies like Virtru

implement ABE in their secu-

rity solutions. Furthermore, several open-source im-

plementations are available (Mosteiro-Sanchez et al.,

2022). Standardization efforts are ongoing, including,

e.g., ETSI’s TS 103 532.

4 DISCUSSION

Based on the descriptions in Section 3.1, it can be

seen that the available computing paradigms address

the requirements set out in Table 1 in highly orthogo-

nal ways, and that some of those requirements are al-

ready effectively addressed by existing technologies.

We provide an overview of the properties of the com-

puting paradigms in Table 2.

For instance, for the underlying trust model, a

https://www.virtru.com/

broad range of options is available, ranging from non-

collusion assumptions in keyless and information-

theoretically secure settings (MPC) over computa-

tional assumptions provided by encryption schemes

with central key material (FHE, FE), up to hardware-

based trust anchors (TEE). Regarding public veriﬁa-

bility and integrity of the computation result, TEEs

and MPC can already provide certain guarantees by

ensuring that computations on sensitive data are ex-

ecuted securely and correctly. For scalability, espe-

cially FL and TEEs provide efﬁcient solutions that

maintain reasonable performance levels, with FL ex-

celling at distributing the computational load to ad-

dress scalability concerns. In addition, FL and TEEs

offer considerable ﬂexibility and ease of integration,

making them suitable for diverse applications.

Some requirements cannot yet be fully met by ex-

isting solutions. For example, while FHE and MPC

offer strong privacy guarantees, they struggle with ef-

ﬁciently scaling large datasets due to high compu-

tational and communication overheads. These tech-

nologies need further efﬁciency improvements to be

practical for real-time use. Ensuring the veriﬁabil-

ity of computations without compromising privacy

remains a challenge, requiring stronger, more prac-

tical mechanisms for approaches like FE or FL. Al-

though ﬂexible, FL needs more development to han-

dle heterogeneous data sources and complex time se-

ries structures. Most technologies also require skilled

personnel for initial setup. Infrastructure needs are

low when deployed as a service, but can rise signif-

icantly if computations are done locally, increasing

data sovereignty.

Finally, certain requirements are not immediately

addressed by any of the technologies. For instance,

none of the computing paradigms inherently supports

data integrity validation, but they all assume valid in-

put data. For all solutions, a legal analysis must be

carried out on a case-by-case basis, considering fac-

tors such as data criticality or additional safeguards.

Finally, for many of the paradigm-speciﬁc or gen-

eral limitations, additional compatible PPTs exist, as

also discussed in Section 3.2.

Based on the ﬁndings presented in the previous

sections, Figure 1 ﬁnally provides a decision diagram

supporting developers in efﬁciently identifying suit-

able candidates for PPTs in the context of federated

time series analysis. While it is not possible to give

conclusive recommendations for every scenario, we

believe that this diagram can provide valuable sup-

port in identifying a starting point for further investi-

gations and analysis.

CLOSER 2025 - 15th International Conference on Cloud Computing and Services Science

206

Table 2: Comparison of guarantees offered by different computing paradigms.

MPC FHE FE FL TEE

Trust model Non-collusion

assumptions; sup-

port of active and

passive adversary

models

Computations

performed by

untrusted data

processor; conﬁ-

dentiality depends

on proper manage-

ment of secret key

Computations

performed by data

receiver; requires

proper protection

of master secret

key

Trust in data own-

ers to use righteous

data; requires ag-

gregator to protect

raw data

Hardware-based

trust anchor; TEE

provider needs to

be trusted

Scalability Well suited for small to medium data sizes Well suited also for large data sets

Efﬁciency Communicational

costs increase with

the complexity of

the circuit to be

evaluated

Non-interactive.

Computational

costs increase with

complexity of the

circuit

Non-interactive.

Costs depend on

computation to be

computed and sce-

nario (input from

one or multiple

parties, etc.)

Computational

costs decreased

through local

lightweight models

but increased com-

municational costs

Moderate overhead

compared to plain-

text computation

Flexibility Support of arbitrary functions;

complexity increases with multi-

plicative depth of circuit

Limited, as

type/class of

computations need

to be known up-

front

Support of a broad

range of FL meth-

ods; increased

communicational

costs with in-

creased complexity

of functions

Support of arbi-

trary functions;

complex adminis-

trative processes by

TEE provider

Data integrity None by default, cf. Section 3.2

Veriﬁability Resilience against

a certain number

of malicious nodes

can be achieved.

Publicy veriﬁable

protocols exist.

Not by default, but

speciﬁc veriﬁable

FHE schemes exist

None by default,

yet computation

can be carried out

be data receiver,

reducing the need

for veriﬁability

None by default,

but impact of data

quality traceable

Integrity based on

the trust assump-

tions into the se-

cure enclave

Ease of integra-

tion

If hosted on-

premise, comput-

ing infrastructure

and trained person-

nel required.

Low integration

costs if used as-

a-service, initial

setup may require

trained personnel.

Key manage-

ment may require

organizational

processes.

Low to moderate.

Key manage-

ment may require

organizational

processes.

Complex Integra-

tion due to setting

up infrastructure

between nodes

Trained personnel

needed in develop-

ment phase; com-

plex organizational

processes to certify

code

Regulatory con-

straints

Requires analysis on a case-by-case basis

5 CONCLUSION

This research highlights the critical need for robust

PPTs in the federated analysis of sensitive time-series

data across different domains. Each considered do-

main presents unique challenges and requirements.

Our evaluation of advanced PPTs and selected exten-

sions reveals their respective strengths and limitations

in meeting these requirements. Furthermore, given

the complex state of affairs for PPTs for time series,

we provide developers with guidance to identify can-

didate PPTs for speciﬁc scenarios. We consider this

an important step towards making those tools accessi-

ble also to engineers. At the same time we acknowl-

edge the need for further evaluation and higher gran-

ularity, which may be achieved in close collaboration

with stakeholders from different domains.

We recognize several research gaps, including the

need for more efﬁcient algorithms for FHE and MPC

to reduce computational and communication over-

heads. Improving the veriﬁability of computations

in privacy-preserving settings is crucial to ensure that

results can be independently veriﬁed without compro-

Protecting Privacy in Federated Time Series Analysis: A Pragmatic Technology Review for Application Developers

207

Re- evaluate the

need for PPTs

Start

Are the data

coming from

diﬀerent

stakeholders?

Does the

computation

involve sensitive

data?

Is the

computation

known upon data

generation?

Are you aiming for

machine learning?

Does the

computation

involve complex

operations?

Is encrypted data

allowed to leave

your premises?

Is the data size

potentially (very)

large?

Federated

Learning (FL)

Functional

Encryption (FE)

Fully

Homomorphic

Encryption (FHE)

Trusted Execution

Environments

(TEEs)

Secure Multi- Party

Computation

(MPC)

Are authenticity

guarantees on

inputs needed?

Advanced

signatures

Are auditability

and/or

veriﬁability

needed?

Zero-

Knowledge

Proofs

Do you have

computation

infrastructure

available?

No suitable PPT

found

Do you have

computation

infrastructure

available?

Is strong access

control on outputs

needed?

Advanced

encryption

YES

YES/NO

Computing

Paradigms

Extensions

LEGEND

Are you

outsourcing your

time series

analysis?

Figure 1: Decision tree supporting the selection of appro-

priate PPTs for federated time series analysis.

mising privacy. Improving the interoperability of dif-

ferent PPTs is essential to facilitate seamless integra-

tion and transition between methods as required by

speciﬁc use cases. In addition, the development of

optimized solutions tailored to the speciﬁc needs of

each domain, especially when dealing with time se-

ries data, is of paramount importance.

Future research should focus on algorithmic in-

novation, efﬁciency benchmarking, cross-technology

integration, domain-speciﬁc solutions, and regulatory

alignment. Investment in the development of new

algorithms and protocols to improve FHE and MPC

efﬁciency and scalability is crucial. Benchmarking

frameworks to evaluate PPT performance in real-

world scenarios will provide additional insights. Inte-

grating multiple PPTs to leverage strengths and mit-

igate weaknesses will enhance their practical appli-

cability. Tailored solutions for speciﬁc requirements

will be needed to address domain-speciﬁc needs.

Engaging with regulators to ensure alignment with

evolving data protection regulations will be essential

for adoption.

Addressing these research gaps and pursuing the

proposed agenda will lead to signiﬁcant progress in

privacy-preserving time series analysis. Ultimately,

these advances will enable the secure and efﬁcient

analysis of sensitive data across critical domains.

ACKNOWLEDGEMENTS

This work was funded by the Austrian Research Pro-

motion Agency FFG within the PRESENT project

(grant no. 899544), and the CHIST-ERA project RE-

MINDER through the Austrian Science Fund (FWF):

I 6650-N. Views and opinions expressed are however

those of the authors only and do not necessarily reﬂect

those of the funding agency. Neither the FFG nor the

granting authority can be held responsible for them.

REFERENCES

Bachlechner, D., La Fors, K., and Sears, A. M. (2018). The

role of privacy-preserving technologies in the age of

big data. In WISP 2018.

Blaze, M., Bleumer, G., and Strauss, M. (1998). Divertible

protocols and atomic proxy cryptography. In Nyberg,

K., editor, EUROCRYPT 1998, volume 1403 of LNCS,

pages 127–144. Springer.

Bogetoft, P., Christensen, D. L., Damgård, I., Geisler, M.,

Jakobsen, T. P., Krøigaard, M., Nielsen, J. D., Nielsen,

J. B., Nielsen, K., Pagter, J., Schwartzbach, M. I., and

Toft, T. (2009). Secure multiparty computation goes

live. In Dingledine, R. and Golle, P., editors, FC 2009,

volume 5628 of LNCS, pages 325–343. Springer.

Bonawitz, K. A., Ivanov, V., Kreuter, B., Marcedone, A.,

McMahan, H. B., Patel, S., Ramage, D., Segal, A.,

and Seth, K. (2017). Practical secure aggregation for

privacy-preserving machine learning. In Thuraising-

ham, B., Evans, D., Malkin, T., and Xu, D., editors,

ACM CCS 2017, pages 1175–1191. ACM.

Boneh, D., Sahai, A., and Waters, B. (2011). Functional

encryption: Deﬁnitions and challenges. In Ishai, Y.,

editor, TCC 2011, volume 6597 of LNCS, pages 253–

273. Springer.

Brakerski, Z. (2012). Fully homomorphic encryption with-

out modulus switching from classical gapsvp. In

Safavi-Naini, R. and Canetti, R., editors, CRYPTO

2012, volume 7417 of LNCS, pages 868–886.

Springer.

Brockwell, P. J. and Davis, R. A. (2016). Introduction to

Time Series and Forecasting. Springer.

CLOSER 2025 - 15th International Conference on Cloud Computing and Services Science

208

Chaum, D. and van Heyst, E. (1991). Group signatures.

In Davies, D. W., editor, EUROCRYPT 1991, volume

547 of LNCS, pages 257–265. Springer.

Chen, G. and Zhang, Y. (2023). Securing tees with veri-

ﬁable execution contracts. IEEE TDSC, 20(4):3222–

3237.

Chiang, H.-L., Chia, Y.-Y., Linand, H.-S., and Chen, C.-

H. (2016). The Implications of Tobacco Smoking

on Acute Postoperative Pain: A Prospective Obser-

vational Study. Pain Research and Management,

2016:9432493.

da Gama, M. B., Cartlidge, J., Polychroniadou, A., Smart,

N. P., and Alaoui, Y. T. (2022). Kicking-the-bucket:

Fast privacy-preserving trading using buckets. In

Eyal, I. and Garay, J. A., editors, FC 2022, volume

13411 of LNCS, pages 20–37. Springer.

Evans, D., Kolesnikov, V., and Rosulek, M. (2018). A prag-

matic introduction to secure multi-party computation.

Found. Trends Priv. Secur., 2(2-3):70–246.

Gentry, C. (2009). Fully homomorphic encryption using

ideal lattices. In Mitzenmacher, M., editor, STOC

2009, pages 169–178. ACM.

Geppert, T., Deml, S., Sturzenegger, D., and Ebert, N.

(2022). Trusted execution environments: Applications

and organizational challenges. Frontiers Comput. Sci.,

Goldreich, O., Micali, S., and Wigderson, A. (1987). How

to play any mental game or A completeness theorem

for protocols with honest majority. In Aho, A. V., ed-

itor, STOC 1987, pages 218–229. ACM.

Goldwasser, S., Micali, S., and Rackoff, C. (1985). The

knowledge complexity of interactive proof-systems

(extended abstract). In Sedgewick, R., editor, STOC

1985, pages 291–304. ACM.

Groth, J. (2016). On the size of pairing-based non-

interactive arguments. In Fischlin, M. and Coron, J.,

editors, EUROCRYPT 2016, volume 9666 of LNCS,

pages 305–326. Springer.

Helminger, L. and Rechberger, C. (2022). Multi-party com-

putation in the gdpr. In Schiffner, S., Ziegler, S., and

Quesada Rodriguez, A., editors, Privacy Symposium

2022, pages 21–39. Springer.

Jordan, S., Fontaine, C., and Hendricks-Sturrup, R. (2022).

Selecting privacy-enhancing technologies for manag-

ing health data use. Front. Public Health, 10:814163.

Keller, M. (2020). MP-SPDZ: A versatile framework for

multi-party computation. In Ligatti, J., Ou, X., Katz,

J., and Vigna, G., editors, ACM CCS 2020, pages

1575–1590. ACM.

Kiss, P. (2022). Federated Learning of Artiﬁcial Neural

Networks.

Kone

cný, J., McMahan, H. B., Yu, F. X., Richtárik, P.,

Suresh, A. T., and Bacon, D. (2016). Federated learn-

ing: Strategies for improving communication efﬁ-

ciency. CoRR, abs/1610.05492.

Li, J., Li, X., and Zhang, C. (2022). Analysis on security

and privacy-preserving in federated learning. High-

lights in Science, Engineering and Technology, 4:349–

358.

Lind, J., Eyal, I., Pietzuch, P. R., and Sirer, E. G. (2016).

Teechan: Payment channels using trusted execution

environments. CoRR, abs/1612.07766.

Liu, T., Xie, X., and Zhang, Y. (2021). zkcnn: Zero knowl-

edge proofs for convolutional neural network predic-

tions and accuracy. In Kim, Y., Kim, J., Vigna, G., and

Shi, E., editors, ACM CCS 2021, pages 2968–2985.

ACM.

López-Alt, A., Tromer, E., and Vaikuntanathan, V. (2012).

On-the-ﬂy multiparty computation on the cloud via

multikey fully homomorphic encryption. In Karloff,

H. J. and Pitassi, T., editors, STOC 2012, pages 1219–

1234. ACM.

Lorünser, T., Wohner, F., and Krenn, S. (2022a). A privacy-

preserving auction platform with public veriﬁability

for smart manufacturing. In Mori, P., Lenzini, G.,

and Furnell, S., editors, ICISSP 2022, pages 637–647.

SCITEPRESS.

Lorünser, T., Wohner, F., and Krenn, S. (2022b). A ver-

iﬁable multiparty computation solver for the linear

assignment problem: And applications to air trafﬁc

management. In Regazzoni, F. and van Dijk, M., edi-

tors, CCSW 2022, pages 41–51. ACM.

McMahan, B., Moore, E., Ramage, D., Hampson, S., and

y Arcas, B. A. (2017). Communication-efﬁcient learn-

ing of deep networks from decentralized data. In

Singh, A. and Zhu, X. J., editors, AISTATS 2017,

volume 54 of Proceedings of Machine Learning Re-

search, pages 1273–1282. PMLR.

Moshawrab, M., Adda, M., Bouzouane, A., Ibrahim, H.,

and Raad, A. (2023). Reviewing federated learn-

ing aggregation algorithms; strategies, contributions,

limitations and future perspectives. Electronics,

12(10):2287.

Mosteiro-Sanchez, A., Barcelo, M., Astorga, J., and Urbi-

eta, A. (2022). Too many options: A survey of ABE

libraries for developers. CoRR, abs/2209.12742.

Muazu, T., Mao, Y., Muhammad, A. U., Ibrahim, M.,

Kumshe, U. M. M., and Samuel, O. (2024). A fed-

erated learning system with data fusion for health-

care using multi-party computation and additive secret

sharing. Comput. Commun., 216:168–182.

Munjal, K. and Bhatia, R. (2022). A systematic review

of homomorphic encryption and its contributions in

healthcare industry. Complex Intell Systems, 3:1–28.

Panzade, P., Takabi, D., and Cai, Z. (2024). Privacy-

preserving machine learning using functional encryp-

tion: Opportunities and challenges. IEEE Internet

Things J., 11(5):7436–7446.

Rivest, R. L., Shamir, A., and Tauman, Y. (2001). How to

leak a secret. In Boyd, C., editor, ASIACRYPT 2001,

volume 2248 of LNCS, pages 552–565. Springer.

Sabt, M., Achemlal, M., and Bouabdallah, A. (2015).

Trusted execution environment: What it is, and what

it is not. In 2015 IEEE Trustcom/BigDataSE/ISPA,

pages 57–64. IEEE.

Sahai, A. and Waters, B. (2005). Fuzzy identity-based en-

cryption. In Cramer, R., editor, EUROCRYPT 2005,

volume 3494 of LNCS, pages 457–473. Springer.

Protecting Privacy in Federated Time Series Analysis: A Pragmatic Technology Review for Application Developers

209

Shamir, A. (1979). How to share a secret. Commun. ACM,

22(11):612–613.

Shi, Y., Song, H., and Xu, J. (2023). Responsible and effec-

tive federated learning in ﬁnancial services: A com-

prehensive survey. In CDC 2023, pages 4229–4236.

IEEE.

Sidorov, V., Wei, E. Y. F., and Ng, W. K. (2022). Com-

prehensive performance analysis of homomorphic

cryptosystems for practical data processing. CoRR,

abs/2202.02960.

Sielaff, L., Hetﬂeisch, R., and Rader, M. (2023). Evaluation

framework for the use of privacy preserving technolo-

gies for production data. In ICAT 2023, volume 11,

pages 157–163.

Siska, V., Lorünser, T., Krenn, S., and Fabianek, C. (2024).

Integrating secure multiparty computation into data

spaces. In van Steen, M. and Pahl, C., editors,

CLOSER 2024, pages 346–357. SCITEPRESS.

United Nations (2023). United nations guide on privacy-

enhancing technologies for ofﬁcial statistics. United

Nations Committee of Experts on Big Data and Data

Science for Ofﬁcial Statistics, New York.

Viand, A., Knabenhans, C., and Hithnawi, A. (2023).

Veriﬁable fully homomorphic encryption. CoRR,

abs/2301.07041.

Wei, K., Li, J., Ding, M., Ma, C., Yang, H. H., Farokhi,

F., Jin, S., Quek, T. Q. S., and Vincent Poor, H.

(2020). Federated learning with differential privacy:

Algorithms and performance analysis. IEEE TIFS,

15:3454–3469.

Yao, A. C. (1982). Protocols for secure computations (ex-

tended abstract). In FOCS 1982, pages 160–164.

IEEE Computer Society.

Zhang, M., Wei, E., and Berry, R. (2021). Faithful edge fed-

erated learning: Scalability and privacy. IEEE Journal

on Selected Areas in Communications, 39(12):3790–

3804.

Zhao, Y., Zhou, H., and Wan, Z. (2024). Superﬂ: Privacy-

preserving federated learning with efﬁciency and ro-

bustness. IACR Cryptol. ePrint Arch., page 81.

Ziller, A., Trask, A., Lopardo, A., Szymkow, B., Wag-

ner, B., Bluemke, E., Nounahon, J.-M., Passerat-

Palmbach, J., Prakash, K., Rose, N., Ryffel, T., Reza,

Z. N., and Kaissis, G. (2021). PySyft: A Library for

Easy Federated Learning, pages 111–139. Springer.

CLOSER 2025 - 15th International Conference on Cloud Computing and Services Science

210